Losses

Source: loss/Losses.luau

Additional losses built from Tensor primitives (Tensor is passed via deps). All return 0-dim scalar Tensors ready for backward(). crossEntropy(logits, targets, smooth?) numeric-stable, mean over rows logSoftmax(logits) -> [N, V] klDiv(logitsP, logitsQ) mean row-wise KL(P || Q) jsDiv(logitsP, logitsQ)? (identity-aggregated P & Q means) focalLoss(logits, targets, gamma?, alpha?, smooth?) binaryCrossEntropy(logits, targets01) (logits = unscaled scores)

Methods

logSoftmax(logits: any)

klDiv(logitsP: any, logitsQ: any)

jsDiv(logitsP: any, logitsQ: any)

focalLoss(logits: any, targets: { number }, gamma: number?, alpha: any?)

binaryCrossEntropy(logits: any, targets: any)

lmCrossEntropy(logits: any, inputs: any, smooth: number?)